이미지: METAL LAB 생성
Summary
- Anthropic updated the biology-related safeguards in Claude Fable 5
- Testing showed biology-related fallback rates dropped by about 85%
- The company said the model can now handle a wider range of everyday health and education questions
- 대상 모델
- Claude Fable 5
- 업데이트 내용
- 생물학 관련 안전장치(세이프가드) 조정
- 효과
- 생물학 관련 폴백 약 85% 감소(전 제품 표면 기준)
- 목적
- 오탐(false positive) 감소, 일상 건강·교육 질문 지원 확대
- 발표 경로
- X(구 트위터) 공식 Claude 계정
- 발표일
- 2026-08-07
What changed
Anthropic announced via its official X account that it has updated the biology-related safeguards in its Claude Fable 5 model. The goal of the adjustment is to reduce "false positives" — cases where questions that pose no real risk are nonetheless blocked. The company said that during testing, the rate of biology-related refusals, or fallbacks, dropped by about 85% across all product surfaces.
The false-positive reduction figure
The only figure disclosed is the fallback reduction rate, but the scale of the drop is considered significant. Below are the key changes confirmed in the announcement.
| Item | Detail |
|---|---|
| Biology-related fallback reduction | About 85% bar:85 |
| Scope of application | All product surfaces |
| Target use cases | Everyday health and education questions |
However, the baseline fallback frequency used to calculate this reduction rate and the specific testing methodology were not disclosed in the original announcement, leaving some uncertainty.
Background: why the safeguards were adjusted
Biology-related safeguards are designed to filter out requests for dangerous biological information, but when set too sensitively, they can also block health information or educational questions. Anthropic said it adjusted Claude Fable 5 to reduce this kind of false positive. In its official announcement, the company explained, "Fable can now assist on a wider range of everyday health and educational questions."
Previous coverage of AI model safeguards and false-positive issues can also be found in METAL LAB's AI safety coverage.
Remaining questions
The announcement was brief, and details such as the scale of testing, the prior version used for comparison, and separate verification of whether safety was compromised were not disclosed. Whether the ability to block genuinely dangerous requests remains intact after relaxing the biology safeguards is a point that still needs further confirmation, and it remains to be seen whether the company will follow up with additional explanation.
